manuscript-argument-compass

Manuscript Argument Compass

Manuscript Argument Compass is a reusable Agent Skill for keeping one evidence-bounded argument coherent across an academic manuscript.

The included skill, steer-question-driven-manuscript, operates at the meta-writing level. It identifies the paper's latent governing inquiry and internal governing proposition, checks paper identity across sections and displays, maps claims to evidence, and hands a stable argument compass to specialised drafting or reporting skills.

It does not replace domain-specific writing, reporting, statistics, citation, or venue guidance.

What it helps with

• diagnosing drift between Introduction, Methods, Results, Discussion, and Conclusion;
• separating the paper's primary argument from attractive secondary branches;
• checking whether the design can adjudicate the intended claims;
• organising Results around the governing inquiry rather than experiment chronology;
• auditing the reviewer-first path through Title, Abstract, displays, Conclusion, Methods, and Introduction;
• designing figures and tables as evidence-bearing argument units;
• keeping alternative explanations distinct from scope or design limitations;
• calibrating claim strength without making the manuscript defensive.

Repository layout

skills/
└── steer-question-driven-manuscript/
    ├── SKILL.md
    ├── agents/
    │   └── openai.yaml
    └── references/
        ├── introduction-argument.md
        ├── venue-modes.md
        └── evidence-integrity.md

The skill content has one canonical source. Codex and Claude Code use the same SKILL.md and references; only installation and invocation differ.

Install for Codex

Copy skills/steer-question-driven-manuscript/ into:

• $CODEX_HOME/skills/ when CODEX_HOME is configured; or
• ~/.codex/skills/ otherwise.

Start a new task after first installation. Invoke it explicitly with:

Use $steer-question-driven-manuscript to identify this paper's governing
inquiry, run the Level 0 identity checksum, and prepare a handoff for the most
appropriate academic-writing skill.

The agents/openai.yaml file provides Codex/OpenAI interface metadata.

Install for Claude Code

For a personal skill available across projects, copy the same directory into:

~/.claude/skills/steer-question-driven-manuscript/

For a project-scoped installation, copy it into:

.claude/skills/steer-question-driven-manuscript/

Invoke it with:

/steer-question-driven-manuscript Audit this manuscript's governing
proposition, evidence chain, visual argument, and section-level drift.

Claude Code reads SKILL.md and its referenced resources. It does not require the Codex-specific agents/openai.yaml.

Suggested workflow

1. Supply the manuscript or available artifacts, target venue, and article type.
2. Let the skill reconstruct the latent inquiry rather than forcing a polished research-question sentence.
3. Review the governing proposition and Level 0 checksum.
4. Resolve evidence gaps or paper-identity mismatches before sentence editing.
5. Hand the compass to a specialised writing skill.
6. Re-audit the finished draft for argument drift and unsupported upgrades.

Scope and evidence boundary

The skill can organise supplied evidence and expose contradictions. It must not invent facts, citations, methods, results, uncertainty estimates, or generalisability. Authors remain responsible for scientific verification and the final manuscript.

Acknowledgement

The initial motivation for the Introduction component was influenced by Shih-Hsin Wang's article, 如何寫出一個有說服力的 Introduction. The present repository develops an independently worded, manuscript-wide framework. See NOTICE.md for provenance details.

License

Original material in this repository is licensed under the Creative Commons Attribution 4.0 International License.